ajustement to avoid oom err

fix "train_on_responses_only not set correctly"

bugfix: ValueError: Unsloth: instruction_part and response_part must be given!

Re-apply unsloth following official guidiance

fix for AttributeError

import automodel for inference test

disable evaluation to save ram as it did crash on v100 at eval phase

rename wandb project name for train with unsloth .py

Add padding and truncation on examples to fix max length

tell trainer to mask user tokens to avoid llm to slef-talks during inference

refs #30267, Add perplexity-based validation in unsloth-based train script

refs #30267, fine-tune with around 20M tokens subset of TaiwanChat

Add script to efficiently train with unsloth